4.5 stars, and half the internet has never heard of it

Nobody hired us to do this either

Same rules as before. No relationship with the hotel, no client engagement, no inside access. Every figure below comes from public listings that anyone can open in a browser, collected in a single scan on 30 July 2026. You can read the whole thing yourself: the reputation audit is public, every platform and every finding.

This is the third lens we have pointed at the same building. Site Pulse crawls the site's technical layer, and gave it 62 out of 100. AI Vision reads the site the way a model would, and gave it 82. Both of those are about what the hotel publishes. A reputation audit is about what everyone else publishes, which the hotel controls far less and worries about far more.

Where it shows up, it is excellent

Start with the part that is unambiguously good. On Google, The Shelbourne holds 4.6 stars from 3,917 reviews, and the shape of those reviews is healthier than the average alone suggests. Three quarters of them are five stars. Fewer than one in twenty is a one or a two.

That distribution is worth more than the headline number. A 4.6 built from a wide spread of threes and fours is a hotel that is reliably fine. A 4.6 built from 77% fives and a thin tail of ones is a hotel that is usually excellent and occasionally has a bad night. The second is a much stronger position, and it is invisible if you only ever look at the average.

The other two correct listings agree. TripAdvisor shows 4.5 from 5,401 reviews. Facebook shows 4.6 from 1,185. Three independent audiences, three effectively identical verdicts, across ten and a half thousand reviews. There is no reputation problem here in the ordinary sense of the phrase.

One caveat on our own method: the TripAdvisor figure is read from a Google rich snippet rather than from TripAdvisor's own API, so it is as fresh as Google's cache of that page. The Google Business numbers, including the distribution above, come straight from the profile.

Everyone praises the staff, everyone mentions the rooms

Scores are what a dashboard reads. Sentences are what a language model reads. When an assistant answers "is The Shelbourne worth it", it summarises prose rather than averaging stars, so the prose is where the answer gets made. Across the ten most recent Google reviews the audit captured, the same two threads run through the good and the bad alike.

The staff are the hotel's strongest asset, and reviewers name them individually:

The rooms are the recurring complaint, and they turn up inside five-star reviews as well as the one-star ones. Note that even the review above, a full five stars, opens the subject unprompted. The harshest review in the sample makes the same split explicitly:

A third theme is doing real work for them: afternoon tea. Three of the ten recent reviews are about the tea rather than a stay at all, one calling it the best in Dublin. That is a distinct product with a distinct audience showing up in the same review stream as the hotel, and it is the sort of thing a hotel usually discovers by accident.

This matters commercially because it is extractable. Any assistant asked to compare Dublin hotels can build the sentence "grand building, exceptional staff, rooms that reviewers describe as small" out of this corpus without inventing anything. That summary is going to be given to prospective guests whether or not the hotel likes it. Fighting the reviews is a waste of a quarter. The remedy is to make the hotel's own pages describe the rooms accurately enough that the two sources agree.

Seven of the last ten reviews got no answer

Of the ten most recent Google reviews, three carry a reply from the business. That is a 30% response rate, and it is the second-largest fixable item the audit flagged.

Which three got answered is the more interesting part. All three replies are on five-star reviews. The one-star review about the rooms, the three-star review about a tour group's stay and the two-star review about afternoon tea are all sitting there unanswered. That is the exact inversion of what review management is for. A reply to a happy guest is a courtesy; a reply to an unhappy one is the only chance the hotel gets to put its version of events into the same page a model will later read.

Be honest about the sample: ten reviews is what this audit captures, not the full history, so 30% is a reading of recent behaviour rather than a lifetime statistic. The pattern of which ones got replies is what we would act on, not the percentage.

Two of the five listings are different buildings

The audit reported that it found the brand on five of the ten platforms it checked, and averaged 4.54 stars across them. We went through those five by hand, and two of them are not The Shelbourne in Dublin.

Both false matches are obvious once you look. Booking.com's own URL puts the property in Japan and its listing name carries the characters for Furano City. Expedia's URL begins with the destination, and the destination is Liverpool. The Japanese hotel has four reviews. The Liverpool apartments have eighty and a perfect five stars, which is what eighty reviews and no bad night looks like.

This is our miss, and we will say plainly what it cost. The 4.54 average is arithmetically correct - 4.6, 4.5, 4.6, 4.0 and 5.0 average to exactly 4.54 - but two of its five inputs are other buildings, and they happen to pull in opposite directions, so the error mostly cancels out and looks harmless. Strip them and the three real listings average 4.57. Coverage is where the damage is: not five platforms out of ten but three, and the composite score is built on the wrong denominator. We ship a "wrong listing?" control on reputation audits for exactly this, and on this run nobody had used it.

The wider point is not about our scraper. A brand-name search for "The Shelbourne" is ambiguous on its own terms - there is a hotel in Hokkaido, an apartment block in Liverpool, a Shelbourne Park greyhound stadium a few kilometres from the hotel, and a Dublin football club. Anything doing entity resolution from a name and a rating faces the same problem we did, and assistants do this constantly. The hotel's defence is the machine-readable identity layer, which is precisely the thing our site audit found thin: no Wikidata entity match at all, and a sameAs block listing four social profiles and nothing authoritative. A strong entity record is how a 200-year-old landmark stops being confused with an apartment block in Liverpool.

The template asked a hotel to be on G2

The false matches are not the only place this run flattered itself. The audit checked ten platforms, and three of them were G2, Capterra and Glassdoor. Two of those are software review sites. The third is an employer review site. A Dublin hotel has no business having a G2 profile, and the audit still counted all three as gaps, each one dragging the coverage figure down.

That is a template configuration, not a bug, and it is ours to explain better. The scoring instruction the tool generated from it - "expanding your presence on platforms like Trustpilot, Yelp, G2, Capterra and Glassdoor will significantly boost your online visibility" - is bad advice for this business, and it scored 80 out of 100 for priority. Told to check ten platforms, the tool checked ten platforms. Nobody told it that four of them were the wrong question.

That leaves Trustpilot and Yelp as the only defensible gaps, and even those are arguable for a luxury hotel whose guests review on TripAdvisor and Google. Trustpilot is the one absence we would act on, because a booking-heavy business benefits from a presence there and The Shelbourne has none.

The lesson generalises past our tool. Any reputation score is a measurement against a list of platforms somebody chose. Read the list before you read the score, and take the platforms off it that your customers do not use. Otherwise you will spend a quarter chasing a number that only measures how odd your template was.

What we would actually do here

Search your own brand name the way a machine would. Not on Google, where you rank first and everything looks fine, but on each booking and review platform in turn. If a differently-located business with your name comes back, you have an entity problem that no amount of review-gathering will fix.

Reply to the bad ones first. The Shelbourne's three replies are all on five-star reviews. Reverse that. An unanswered one-star review is the only version of that incident on the record, and it is on the page an assistant summarises.

Claim a Wikidata entity. It is free, it takes an afternoon, and it is the closest thing there is to a machine-readable identity document. For a business whose name collides with three others, it is the single highest-value item on this list.

Prune your platform list before you chase coverage. Three of the ten platforms in this audit were irrelevant to a hotel. Coverage against a list that includes them is a worse number than coverage against a list that does not, and chasing it wastes real money.

Read the distribution, not the average. 77% five stars with a thin one-star tail is a different business from a 4.6 built out of threes and fours, and it should be managed differently.

Run a reputation audit on your own brand

Questions this usually raises